Uselessly spends time materializing tensor proto that will never be used

[ezyang]

In backend.py:

        # Caffe2 predictor requires all input blobs (including the
        # real model inputs) are initialized in init_net
        for value_info in graph_def.input:
            if value_info.name in initialized:
                continue
            op_def = caffe2_pb2.OperatorDef()
            op_def.output.extend([value_info.name])
            op_def.type = 'GivenTensorFill'

            shape = list(d.dim_value for d in value_info.type.tensor_type.shape.dim)
            # TODO: Putting this in the init net will make it run faster, but it
            # causes some tests to fail...
            # shape = (1,)

            shape_arg = op_def.arg.add()
            shape_arg.name = 'shape'
            shape_arg.ints.extend(shape)

            values_arg = op_def.arg.add()
            values_arg.name = 'values'
            values_arg.floats.extend(np.ones(shape).flatten().tolist())

            init_net.op.extend([op_def])

This is pretty pointless (we never actually use the values from the init net) AND it's really expensive (because we materialize a tensor proto for the inputs.) The shape hack used to work by Caffe2 rejects it now.